Closed C4712 closed 3 years ago
What's wrong?
GPU execution requested, but not compiled with GPU support,this problem maybe because wrap_ctc can't load gpu,but i replace by torch.nn.Ctcloss,the loss alway in 6 or 7,the cer is 1.0,
because in server write code it is not essay to find the wrap_ctc source code,can you help me solve the problem
@5859wang See this document: https://github.com/yeyupiaoling/MASR/edit/master/README.md
c10/cuda/CUDAGuard.h: 没有那个文件或目录
错误,将pytorch_binding/src/binding.cpp
将#include <c10/cuda/CUDAGuard.h>
修改成#include "ATen/cuda/CUDAGuard.h"
。
git clone https://github.com/SeanNaren/warp-ctc.git
cd warp-ctc
mkdir build
cd build
cmake ..
make -j4
cd ../pytorch_binding
python setup.py install
i already install wrap_ctc and can successed exexutor the wrap_ctc test_gpu.py file,
@5859wang What's your versionof CUDA?warp-CTC only support CUDA10.1.
I don't understand you said but not compiled with GPU support
my cuda version is 10.2,but i can run the wrap-ctc test_gpu.py fairseqe) [root@localhost tests]# python test_gpu.py =========================================================================== test session starts =========================================================================== platform linux -- Python 3.6.2, pytest-6.1.2, py-1.9.0, pluggy-0.13.1 rootdir: /data/wh/MASR/warp-ctc/pytorch_binding, configfile: setup.cfg plugins: hydra-core-1.0.4 collected 5 items
test_gpu.py ..... [100%]
============================================================================ 5 passed in 5.76s
your mean is the wrap-ctc install successed,but can't transfer cuda?only cuda10.1 can run the Ctcloss,cuda10.2 install successed can't still transfer
@5859wang No, CUDA 10.2 can also be used, and mine is cuda10.2.May I see the error message?
GPU execution requested, but not compiled with GPU support